Dynamic aodv


Published on

Published in: Education, Technology
1 Like
  • Be the first to comment

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide

Dynamic aodv

  1. 1. IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 5, JUNE 2009 2471 A Dynamic Anomaly Detection Scheme for AODV-Based Mobile Ad Hoc Networks Hidehisa Nakayama, Member, IEEE, Satoshi Kurosawa, Abbas Jamalipour, Fellow, IEEE, Yoshiaki Nemoto, Senior Member, IEEE, and Nei Kato, Senior Member, IEEE Abstract—Mobile ad hoc networks (MANETs) are usually off meeting networks, disaster and military applications, and formed without any major infrastructure. As a result, they are the entertainment industry. Because the network topology of relatively vulnerable to malicious network attacks, and therefore, MANETs frequently changes, and there is no central manage- security is a more significant issue than infrastructure-based wire- less networks. In MANETs, it is difficult to identify malicious hosts ment entity, all of the routing operations must be performed by as the topology of the network dynamically changes. A malicious the individual nodes in a collaborative fashion. Consequently, host can easily interrupt a route for which it is one of the forming it is unrealistic to introduce an authentication server that can nodes in the communication path. In the literature, there are employ conventional cryptographic schemes to secure the net- several proposals to detect such malicious hosts inside the network. work against attacks from malicious hosts. The typical types of In those methods, a baseline profile, which is defined as per static training data, is usually used to verify the identity and the attacks in MANETs include eavesdropping, address spoofing, topology of the network, thus preventing any malicious host from forged packets, denial of service (DoS), etc. [1]. joining the network. Since the topology of a MANET dynamically Secure routing protocols [2]–[4] in which key-based crypto- changes, the mere use of a static baseline profile is not efficient. graphic technologies [5], [6] are applied have been suggested to In this paper, we propose a new anomaly-detection scheme based meet the increasing demands for MANET security. However, on a dynamic learning process that allows the training data to be updated at particular time intervals. Our dynamic learning besides the topology issue, these methods cannot protect the process involves calculating the projection distances based on mul- network from attacks by a malicious node that has managed tidimensional statistics using weighted coefficients and a forgetting to acquire the network key. Therefore, other security methods curve. We use the network simulator 2 (ns-2) system to conduct that can detect attacks from malicious hosts are required. If a the MANET simulations and consider scenarios for detecting five well-known attack in the TCP/IP protocol stack is launched in types of attacks. The simulation results involving two different networks in size show the effectiveness of the proposed techniques. a MANET, then it is possible to protect the network by using conventional security techniques [7]. However, if the attacker Index Terms—Ad hoc on-demand distance vector (AODV), maliciously uses the specific routing protocol of the MANET, anomaly detection, dynamic learning, forgetting curve, malicious attacks, mobile ad hoc networks (MANETs), projection distance. prevention becomes remarkably difficult [8]. In such a case, it is almost impossible to recognize where and when the malicious node appears. Thus, the attack detection at each node becomes I. I NTRODUCTION necessary [9]. The techniques for detecting the malicious attacks are usu- I NCREASINGLY, mobile ad hoc networks (MANETs) are receiving more attention as part of the next-generation net- work technologies. These networks are usually constructed by ally classified into two categories, namely, misuse detection and anomaly detection. In misuse detection, the method of using mobile and wireless hosts with minimum or no central using a signature-based analysis is widely implemented. In this control point of attachment, such as a base station. These method, the attacks are identified by comparing the input traffic networks can be useful in a variety of applications, such as one- signature with the signatures extracted from the known attacks at the network routers. An anomaly detection is a technique that quantitatively defines the baseline profile of a normal Manuscript received April 16, 2008; revised August 30, 2008. First published system activity, where any deviation from the baseline is treated November 25, 2008; current version published May 11, 2009. The review of this as a possible system anomaly. It is rather easy to detect an paper was coordinated by Prof. X. (S). Shen. attack, the traffic signature of which is identifiable by using H. Nakayama was with the Graduate School of Information Sciences, Tohoku University, Sendai 980-8579, Japan. He is now with the Department misuse detection. However, for those attacks, the type or traffic of Electronics and Intelligent Systems, Tohoku Institute of Technology, Sendai signatures of which are hard to identify by misuse detection, 982-8577, Japan (e-mail: hidehisa@m.ieice.org). the method is rather inadequate. In such cases, those attacks S. Kurosawa was with the Graduate School of Information Sciences, Tohoku University, Sendai 980-8579, Japan. He is now with the Information Technol- can only be detected by using anomaly detection methods. In ogy R&D Center, Mitsubishi Electric Corporation, Kamakura 247-8501, Japan anomaly detection, even when the traffic signature is unknown, (e-mail: Kurosawa.Satoshi@cw.MitsubishiElectric.co.jp). if the baseline profile of a network is delineated a priori, then A. Jamalipour is with the School of Electrical Information Engineering, University of Sydney, Sydney NSW 2006, Australia (e-mail: a.jamalipour@ the abnormality can be recognized. In [10], the effectiveness of ieee.org). such a detection method in wired networks has been demon- Y. Nemoto and N. Kato are with the Graduate School of Information strated. In this method, the baseline profile is preextracted and Sciences, Tohoku University, Sendai 980-8579, Japan (e-mail: nemoto@ nemoto.ecei.tohoku.ac.jp; kato@it.ecei.tohoku.ac.jp). then applied to the same network. However, for MANETs, since Digital Object Identifier 10.1109/TVT.2008.2010049 the network conditions are likely to change, the preextracted 0018-9545/$25.00 © 2009 IEEEAuthorized licensed use limited to: Arulmigu Kalasalingam College of Engineering. Downloaded on August 04,2010 at 09:46:45 UTC from IEEE Xplore. Restrictions apply.
  2. 2. 2472 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 5, JUNE 2009 network state may not correctly represent the state of the current network performance, as well as on the connectivity among the network. This problem indeed influences the accuracy of the nodes in the targeted MANET. anomaly detection method. Deng et al. [29] proposed an approach that requires the Due to the fact that the MANET environment dynamically intermediate nodes to send a route reply (RREP) packet with keeps evolving, envisioning a robust anomaly detection method the next hop information. When a source node receives the becomes imperative to thwart the malicious attacks against it. In RREP packet from an intermediate node, it sends a “Further this paper, we propose a new anomaly detection scheme based Request” packet to the next hop to verify that it has a route on a dynamic learning method. The MANET hosts are mobile to the intermediate node and a route to the destination. As a on their own so that the MANET environment is dynamically response to this request, the intermediate node will send another changing. Our dynamic learning method is based on a statistical RREP packet. When the next hop receives a “Further Request” decision theory that calculates the multidimensional projection packet, it sends a “Further Reply” packet that includes the distance between the current and normal states of the targeted verified result to the source node. Based on the information in host. We propose to use weighted coefficients with a forgetting the “Further Reply” packet, the source node judges the validity curve as its mathematical property has been proved [11], [12] of the route. Again, the method in [30] requires the intermediate to suit our requirements. We conduct network simulations with node to send the route confirmation request (CREQ) to the next five types of attacks in [13]–[15] as a case study that concerns hop node toward the destination, and then, the next hop node one of the most popular MANET routing protocols, i.e., the ad receives the CREQ and looks into its cache for a route to the hoc on-demand distance vector (AODV) [16]. The simulation destination. If it has such a route to the destination, then it sends results of the network simulator 2 (ns-2) [17] demonstrate a route confirmation reply (CREP) message to the source node the effectiveness of the proposed technique, regardless of the with its route information. The source judges whether the path number of nodes in the considered MANET. in RREP is valid by comparing the information with CREP. The remainder of this paper is organized as follows. In In these methods, the routing protocol has to be modified. Section II, we present the problems of conventional detection These modifications may increase the routing overheads, which schemes in attacks against MANETs. In Section III, we present results in the performance degradation of the bandwidth-limited an overview of AODV. Section IV describes the proposed MANETs. detection scheme and the derivation of the essential parameters. In Section V, the simulation results concerning the performance B. Network Monitoring-Based Attack Detection of the proposed scheme are provided. Section VI presents the conclusions and future research scopes. In addition to the aforementioned techniques, an attack detection by network monitoring, which can detect attacks II. R ELATED W ORKS from inside MANETs, has also been proposed. For instance, Kachirski and Guha [31] proposed a method that detects attacks A. Secure Schemes for Routing Procedures by employing distributed mobile agents. Network monitor- Secure ad hoc routing protocols have been proposed as a ing nodes are selected to be able to collect all the packets technique to enhance the security in MANETs. For example, within a cluster, and the decision agents in the nodes are the secure AODV (SAODV) [18], which uses signed rout- used to detect and classify the security violations. The concern ing messages, is proposed to add security to AODV [16]. of this method is that the monitoring nodes will consume A-SAODV [19], [20] is a mild implementation of SAODV that a large amount of energy. Vigna et al. [32] detect attacks uses the RSA [21] as an asymmetric cryptographic algorithm by placing AODV-based State Transition Analysis Technique and the SHA1 [22] as a hash algorithm. The survey conducted (AODVSTAT) sensors within the network and by either ob- by Yih-Chun and Perrig [23] overviewed the various secure serving solely contiguous nodes or trading information with routing protocols and pointed out their drawbacks and advan- other sensors. However, it is necessary to deploy a large number tages. They also proposed a secure on-demand ad hoc network of AODVSTAT sensors on the nodes for detecting a varied routing protocol (Ariadne) [24], which prevents the compro- range of attacks. In addition, a large number of UPDATE mes- mised nodes from tampering with the uncompromised routes, sages may cause an overwhelming congestion in the network. and the secure efficient ad hoc distance (SEAD) [25], which Tseng et al. [33] introduced a method that places a network is a secure routing protocol, using efficient one-way hashing monitor (NM) inside the network. In this method, the NM functions and not using asymmetric cryptographic operations. constantly monitors the packet flow in the network within a In addition, Zhou and Haas proposed a distributed certification certain range to detect any attacks. However, placing effective authority mechanism in which the authentication uses threshold detectors, i.e., mobile agents, sensors, or NMs, is considered to cryptography [2]. In [26], a MANET is divided into clusters, be difficult when the MANET topology dynamically changes. and a certification authority is appointed to each cluster. In [27], One solution to this problem is to observe the packet flow on a method called key predistribution (KPD) scheme is applied. In each node and to detect any potential attack. [28], the authenticated routing for ad hoc networks (ARAN) is proposed by using public-key cryptographic mechanisms based C. Anomaly Detection on the AODV. These methods can only guard against external attacks. However, the internal attacks mounted by the malicious Huang et al. [34] proposed a method in which the packet or compromised hosts may still have a severe impact on the flow is observed at each node. In this method, 141 features thatAuthorized licensed use limited to: Arulmigu Kalasalingam College of Engineering. Downloaded on August 04,2010 at 09:46:45 UTC from IEEE Xplore. Restrictions apply.
  3. 3. NAKAYAMA et al.: DYNAMIC ANOMALY DETECTION SCHEME FOR AODV-BASED MOBILE AD HOC NETWORKS 2473 Fig. 2. Transferring RERR messages on AODV. Fig. 1. Route-discovery process on AODV. every time node S sends an RREQ message. Nodes A and B, are both traffic and topology related are defined. Huang et al. which have received the RREQ message, generate and renew suggested an anomaly detection mechanism with interrela- the route to its previous hop. They also evaluate if this is a tion between features. Moreover, in [35], they constructed an repeated RREQ message and accordingly discard it. If A and extended finite-state automaton (EFSA) according to the B have a valid route to the destination D, then they send an specification of the AODV routing protocol, envisioned nor- RREP message to node S. In the case where the node has no mal condition modeling, and detected attacks with both valid route, they send an RREQ message using broadcasting. specification-based and anomaly-based detection schemes. In The exchange of route information will be repeated until an specification-based detection, the attacks were detected as de- RREQ message reaches node D. When node D receives the viant packets from the conditions defined by EFSA. In addition, RREQ, it sends an RREP message to node S. When node S in anomaly detection, the normal conditions are defined as receives the RREP message, a route is established. In case of the baseline with which the condition of EFSA and also the multiple RREPs received, a node selects an RREP message, amounts of transition statistics are compared. The deviations the Destination Sequence number (Dst_Seq) of which is the from those conditions are then used to detect the potential largest among all the previously received RREPs. However, attacks. For determining the baseline profiles, in both methods, if the Dst_Seqs were the same, then it will select the RREP the training data are extracted beforehand from the same net- message whose hop count is the smallest. work environment where the test data are applied. However, we In Fig. 2, when node B detects a disconnection of route, it note that the MANET topology can rather easily be changed, generates route error (RERR) messages and puts the invalidated and the differences in network states grow larger with time. Fur- address of node D into its list and then sends RERR to node A. thermore, these methods cannot be applied to a network where When node A receives the RERR message, it refers to its route the learning phase has been conducted in another network. Sun et al. [36] proposed an anomaly detection method in map and the current list of RERR messages. If there was a route which mobility is considered. This method computes the recent to the destination for node D included in its map, and the next link change rate (LCRrecent ) and can select the training data, hop in the routing table is a neighboring node B, it invalidates the link change rates of which have the smallest Euclidean the route and sends an RERR message to node S. This way, the distance to LCRrecent . However, the change of network states RERR message can finally be sent to the source node S. can be caused not only by mobility; it may also occur due to the sudden participation and disappearance of nodes in a MANET. B. Classification of Attacks When the nodes in the current MANET differ from those in the training data, the defined baseline profile cannot express the According to the aforementioned features, the malicious current network state. As a result, these methods are rendered nodes can misuse the AODV by forging source IP addresses, inadequate and considered difficult in a MANET environment. destination IP addresses, RREQ IDs, hop counts, Destina- To solve this problem, a normal state needs to be defined by tion Sequence numbers (Dst_Seqs), Source Sequence numbers using the data reflecting the trend of the current situation, and (Src_Seqs), and also by flooding the network with routing pack- this leads to the idea of updating the learning process within a ets. According to prior works (e.g., [13]–[15]), we can classify time interval. By doing so, the attack detection can adaptively the attacks against AODV into routing disruption attacks and be conducted even in a changing network scenario. resource consumption attacks. 1) Routing Disruption Attacks: These attacks interrupt the III. A TTACKS ON AODV P ROTOCOL establishment of a route or destroy an existing route. The most common attacks of this type are the modification of A. Overview of AODV Protocol RREP (same as the Blackhole Attack) and the modifica- The AODV [16] is a reactive routing protocol in which the tion of RREQ. network generates routes at the start of communication. Each 2) Resource Consumption Attack: This attack wastes re- node has its own sequence number, and this number increases sources of a specific node and the network as a whole. The whenever a link changes. According to its sequence number, most common attack of this type is malicious flooding. each node judges whether the channel information is recent. A short explanation of the preceding three attacks is Fig. 1 illustrates the route-discovery process of the AODV. given here. In this figure, node S attempts to establish a connection to 1) Modification of RREP: The Dst_Seq represents the fresh- destination D. First, the source node S refers to the route map ness of routing information in the network. When a at the start of communication. In the case where there is no source node receives multiple RREP messages, it selects route to destination node D, it sends a route request (RREQ) the node that has the largest Dst_Seq value and accord- message by using broadcasting. The RREQ ID increases by one ingly constructs a route. Therefore, a malicious node mayAuthorized licensed use limited to: Arulmigu Kalasalingam College of Engineering. Downloaded on August 04,2010 at 09:46:45 UTC from IEEE Xplore. Restrictions apply.
  4. 4. 2474 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 5, JUNE 2009 intentionally attempt to modify the RREP packet and increase the Dst_Seq value of the RREP message. As a result, a false route will be established, and the legitimate data traffic will be interrupted. In addition, the victim nodes will further spread the false routing information to others, and thus, the damage will propagate throughout the network. In this case, we can consider two types of forged packets. In the first type, the source and destination IP addresses are spoofed or forged to the destination node. In the second type, the destination IP address is forged to the destination node, and the source IP address is spoofed Fig. 3. Feature definition. The traffic features in a time slot are expressed by to a randomly selected node. We call the “Modification of the elements of the p-dimensional vector x. RREP (1)” as an attack that uses the first packet type and we obtain p = 14. According to [10], [34], and [35], a small the “Modification of RREP (2)” as an attack that uses the value for the time slot is preferred, and therefore, Δτ is set to a second packet type. constant small value of 5 s. 2) Modification of RREQ: The RREQ ID represents the Furthermore, in the learning process shown in Fig. 3, a time freshness of an RREQ message in the network. Based interval ΔT is defined. It contains several time slots. In other on the RREQ ID, each node decides whether to forward words, the number of time slots is equal to the number of all an RREQ message. Therefore, a malicious node attempts training samples at a given time interval. Note that if we use a to intentionally increase the RREQ ID when an RREQ shorter time interval ΔT , the data sets contained in one time packet is received. Additionally, when a forged packet interval will decrease. On the other hand, a larger value of ΔT with a false source address in the IP header is sent, the slows down the learning process. route will never be established. The statistics for a time interval define a state in the network, 3) Malicious Flooding: Generally, the RREQ messages are and a further explanation about the statistics is described in broadcasted to select new routes. If a malicious node Section IV-B. sends an excess number of RREQ messages, then the 1) Path Finding Features (Nine Dimensions): The path find- network will become congested with a huge amount of ing features comprise the following: RREQ traffic. In our preliminary experimental results, when a malicious node sends more than 20 RREQ packets 1) number of received RREQ messages (three types); per second, the congestion occurs, which leads to signif- 2) number of forwarded RREQ messages; icant unnecessary delays and packet drops. In this case, 3) number of outbound RREQ messages; we can consider two forged packet types. In the first type, 4) number of outbound RREP messages (two types); the source IP address is forged to a randomly selected 5) number of received RREP messages (two types). node. In the second type, the source IP address is forged For each node, the number of received RREQ messages to a destination node, and the RREQ ID is intentionally includes three types, i.e., messages with their own source IP increased at the same time. We define the “Malicious addresses, messages with their own destination IP addresses, Flooding (1)” as an attack that uses the first packet type and messages with neither source nor destination IP addresses and the “Malicious Flooding (2)” as an attack that uses of their own. When counting the number of received RREP the second packet type. messages, the packets with a matching destination IP address, See [13]–[15] for detailed information on these attacks. source IP address, RREQ ID, or Src_Seq in the training data are recorded once for each time slot. Similarly, the number of outbound RREP messages includes two types, for which the IV. D YNAMIC A NOMALY D ETECTION destination node is itself, and for which it holds the path toward In this section, we first introduce the features that are es- the destination node. The number of received RREP messages sential for our envisioned anomaly detection scheme, and then includes two types: the first type is a packet, both source and delineate the module of the detection scheme based on the destination addresses of which exist in the training data. All projection distances. of the other packets are classified as the second type (with either one or no matching features). As an example, when a node is under attack by the “Malicious flooding,” it receives A. Definition of Features a tremendous amount of RREQ messages, and therefore, the Each node observes its own traffic and uses a time slot to number of received RREQ messages increases. This indicates record the number of packets (messages) according to their the presence of anomaly in the network. types (see Fig. 3). In time slot Δτ , the instantaneous value 2) Path Abnormality Features (Four Dimensions): The path of the network state is expressed by a p-dimension vector abnormality features comprise the following: x = [x1 , x2 , . . . , xp ]T , where each feature xk (k = 1, . . . , p) 1) number of received RERR messages; is measured. In this paper, we define nine features related to 2) number of outbound RERR messages; path finding, four features related to path abnormality, and one 3) number of dropped RREQ messages; feature related to a major characteristic of AODV. Therefore, 4) number of dropped RREP messages.Authorized licensed use limited to: Arulmigu Kalasalingam College of Engineering. Downloaded on August 04,2010 at 09:46:45 UTC from IEEE Xplore. Restrictions apply.
  5. 5. NAKAYAMA et al.: DYNAMIC ANOMALY DETECTION SCHEME FOR AODV-BASED MOBILE AD HOC NETWORKS 2475 Fig. 4. Distance of sample x to the first principal element φi . d(x) is the projection distance. When counting the number of received RREQ messages, the packets with the same destination IP address and Dst_Seq are recorded only once for each time slot. As an example, when a Fig. 5. Example of the division by using the projection distance. node is under attack of the “Modification of RREQ” messages caused by packets with a forged source address in the IP header, From (1) and (2), we use the principal component analysis a large number of RREP messages will not successfully be (PCA) [37] to analyze the statistical nature of the current time able to be sent out. As a result, the number of dropped RREP interval. The PCA is the method that explores the correlations messages will increase, and this acts as a sign of abnormality in between each feature and finds the most important axis to the network. express the scattering of data. Here, the most important axis 3) AODV Characteristic Feature (One Dimension): The denotes the baseline profile of network activity. When an attack AODV characteristic feature comprise the average of the dif- takes place, it generates the deviation sample from this axis. ferences of Dst_Seq in each time slot between the number of By using PCA, the first principal element φi , which reflects the received RREP messages and the one held in the list. approximate distribution of the training data sets, is calculated. When sending or forwarding an RREQ message, each node Here, we consider the projection distance of an input data keeps the destination IP address and the Dst_Seq in its list. sample x as When an RREP message is received, the node looks over the d(x; D i ) = x − xi ¯ 2 − φT (x − xi ). i ¯ (3) list to see if there is the same destination IP address. If it does exist, the difference of Dst_Seq is calculated, and this operation When the projection distance is larger than the threshold MI , is executed for every received RREP message. The average of it is evaluated as this difference is finally calculated for each time slot as the feature. Due to the link error in the ad hoc networks, sometimes d(x; D i ) > MI : attack (4) the nodes might receive an old RREP message. In this case, d(x; D i ) ≤ MI : normal. the newly received Dst_Seq in RREP is smaller than the one Here, when Mi is the maximum value of projection distance already kept in the list. When this happens, the calculation for node i in the training data sets D i , the suffix I of MI is is excluded. In a normal state, the Dst_Seq increases in a extracted from all the nodes (N ) as relatively stable pace. On the contrary, when a node is receiving “Modification of RREP” attacks, this value drastically changes, I = arg max Mi i i=1,...,N and thus, we may recognize this particular abnormality. where B. Detection Module by Projection Distance Mi = max d(x; D i ). (5) x∈D i In pattern recognition, based on statistical decision theory, Fig. 5 shows a rough image of determining the normal or at- the distance measure is an effective way to formulate the differ- tack states by using the projection distances in two dimensions. ent types of categories because the same category is distributed in the close area in a multidimensional feature space [37]. Here, the normal and attack states as two different categories can be C. Proposal of Dynamic Anomaly Detection considered. In this section, we describe the detection module by Since the network topology easily changes in MANET, the using the projection distance (see Fig. 4). current state may not appropriately be expressed over time. Let us consider a training data set D i = {x} collected by Therefore, by only using the method described in Section IV-B each node i (i = 1, . . . , N ), where N is the number of all nodes to define the normal state, it is rather insufficient to reflect the participating in MANET, and the current time interval consists changing situation of MANET, and a learning method that can of Di time slots (Di = |D i |, in case of using all training follow these changes is indispensable. We explain the idea of samples). First, we calculate the mean vector and the covariance dynamically updating the training data sets in the remainder of matrix at node i as this section. 1 Let T0 be the current time interval, and let T1 be the first ¯ xi = x. (1) Di time interval. By using the data collected in T1 , initially, the x∈D i first principal element is calculated, and then the calculated first 1 principal element is used in the following time interval T0 for Σi = (x − xi )(x − xi )T . ¯ ¯ (2) Di anomaly detection. If the state in T0 is judged as normal, then x∈D iAuthorized licensed use limited to: Arulmigu Kalasalingam College of Engineering. Downloaded on August 04,2010 at 09:46:45 UTC from IEEE Xplore. Restrictions apply.
  6. 6. 2476 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 5, JUNE 2009 Here, we consider all the training data sets U i = {D i (1) ∪ D i (2) ∪ · · · ∪ D i (mi )} . (10) By using PCA, the first principal element φi (0) is calculated. As a general scheme, the distance d(x; U i ) of the input data sample x can be computed from (3) and then evaluated as d(x; U i ) > MI : attack (11) d(x; U i ) ≤ MI : normal. Here, when Mi is a maximum value of the projection distance for node i in all the training data sets U i , the suffix I of MI is extracted from all the nodes (N ) as Fig. 6. Flow chart of the proposed method for learning and evaluation. I = arg max Mi the corresponding data set will be used as the training data set. i i=1,...,N Otherwise, it will be treated as the data including attack, and it where will consequently be discarded. This way, we keep on learning the normal states of the network. This procedure is shown Mi = max d (x(0); U i ) . (12) in Fig. 6. x(0)∈U i As mentioned earlier, when updating the database, it is possible to use the most recent data set. However, since the most recent data set is easily affected by the sudden change D. Derivation Algorithm of Parameters in the network, it is necessary to take the time series model into According to [38], the mobility metric of the MANETs is consideration to keep the database from being too sensitive to expressed by using the number of neighbor nodes. Using the the changes in the network topology. Here, we use the forgetting number of neighbors, the number of training data sets mi curve [11], [12] as the weighting function to adjust the degree used in the learning process and the parameter ai in (6) can of importance of the time slot. The forgetting curve aims at dynamically be determined. Assume that for a given node i, at reducing the weight when the data become old and of less time t, its neighbor set is S i (t), t = 0, 1, 2, . . . , mi , mi+1 . If significance. S i (0) ∩ S i (mi + 1) = ∅, then we can recognize that the net- Suppose using mi data sets as the training data. Fig. 7 work state has considerably changed, and then mi is determined shows how to weigh the data sets while learning. λi (t), t = as the number of training data sets. Next, we consider ai in (6). 1, 2, . . . , mi , are the forgetting coefficients that correspond to ai represents the change in the size of the considered network. each training data set, respectively. As shown in Fig. 8, the The change in size of a network is expressed by the change in forgetting curve is expressed as the number of its neighboring nodes. Assume that for a given λi (t) = λi (0)e−ai Tt node, at the first time interval (t = 1), its neighbor set is S i (1). = λi (0) · exp(−ai ΔT · t), i = 1, . . . , N (6) |S i (0) − S i (1)| is the number of new neighbors during ΔT , and |S i (1) − S i (0)| is the number of neighbors that moved where the current coefficient λi (0) and the common time away. Then, ai can be calculated as interval ΔT are constants. λi (t) are constrained by mi |S i (0) − S i (1)| |S i (1) − S i (0)| ai = + . (13) 1= λi (t), i = 1, . . . , N. (7) N N t=1 Here, ai is normalized by N (the number of all nodes partici- Additionally, we describe the derivation algorithm of determin- pating in MANET). Next, we give an example of the simulation ing the parameters mi , ai in Section IV-D. Using the number of data. Fig. 9 shows the changes of the number of training data data sets mi and the forgetting coefficients λi (t), the statistics sets mi . We can see that mi dynamically varies as the time of the current state “0” can be calculated from (1) and (2) as elapsed. mi ¯ xi (0) = λi (t)¯ i (t) x V. P ERFORMANCE E VALUATION t=1 mi λi (t) In this section, we describe the details to evaluate the pro- = x(t) (8) posed method. t=1 Di (t) x(t)∈D i mi Σi (0) = λi (t)Σi (t) A. Simulation Environment t=1 mi The experiments were carried out by using ns-2 (ver. 2.27) λi (t) = (x(t)− xi (t)) (x(t)− xi (t))T . (9) ¯ ¯ [17]. We assume that the simulation network being used is in a t=1 Di (t) place where various events in a MANET can occur [39], [40]. In x(t)∈D iAuthorized licensed use limited to: Arulmigu Kalasalingam College of Engineering. Downloaded on August 04,2010 at 09:46:45 UTC from IEEE Xplore. Restrictions apply.
  7. 7. NAKAYAMA et al.: DYNAMIC ANOMALY DETECTION SCHEME FOR AODV-BASED MOBILE AD HOC NETWORKS 2477 Fig. 7. Renewing training data using the forgetting coefficients. Fig. 8. Ebbinghaus’ forgetting curve. Fig. 10. Mobility pattern for RWP in 5 s. between 0 and 5 m/s. The pause time was set to 10, 50, 100, 200, and 500 s, respectively. For example, Fig. 10 shows the moving pattern within 5 s. To start the learning process, the first normal state, which excludes the attack data, was manually preextracted from the training data. This is because our proposed method detects the possibility of attacks according to the degree of which a state deviates from the normal state. Here, the first time interval is set to 300 s. This is a period in which enough normal state samples can be collected. We also deemed that it necessary Fig. 9. Example of the number of training data sets, dynamically updated. to shorten the updating interval as the mobility rates increase. However, the shorter the updating interval, the more processing this simulation, the run time is 10 000 s, and five types of attacks overhead is required. Therefore, more battery power will be were randomly executed from 2500 to 5000 s. All of the nodes, consumed. From these facts, it is necessary to take into account except the attack node, employed the proposed method to detect the MANET environment and the available battery power to attacks. The simulations were performed for the following two determine the time interval of updating. In our results, the time scenarios: 1) a 50-node network with a network topology of interval of updating ΔT was set to 600 s. 1000 m × 1000 m and 2) a 100-node network with a network topology of 2000 m × 2000 m. The traffic loads were constant bit rate flows with a data packet size of 512 B. In 1), the load B. Simulation Results was varied by using 40 flows (at four packets per second). In 2), the load was varied by using 80 flows (at four packets per To evaluate our proposed methods, we assume the following second). The 802.11 Media Access Control (MAC) layer was three ways of using the training data sets: used with a transmission range of 250 m, and it was set for a 1) M-1: method of using only the initial training data set; 2-Mb/s throughput. As for the moving pattern for each node, we 2) M-2: method of using the most recent training data set at use a random waypoint (RWP) model [41] in which each node every time interval; randomly selects the destinations in the designated simulation 3) M-3: method of using the training data sets dynamically area with random speeds. Here, the node velocity was set decided at every time interval.Authorized licensed use limited to: Arulmigu Kalasalingam College of Engineering. Downloaded on August 04,2010 at 09:46:45 UTC from IEEE Xplore. Restrictions apply.
  8. 8. 2478 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 5, JUNE 2009 Fig. 11. Detection module by projection distance against modification of RREP(1), N = 50. (a) M-1 (conventional method). (b) M-2 (reference method). (c) M-3 (proposed method). Fig. 12. Detection module by projection distance against modification of RREP(1), N = 100. (a) M-1 (conventional method). (b) M-2 (reference method). (c) M-3 (proposed method). “M-1” is the conventional method, and “M-3” is our proposed rapid changes of network topology, this point is incorrectly method. For reference, we also show the result of “M-2,” which evaluated and produces a false positive. As a whole, we can is the simplest case of our proposed method. see that, in the proposed method, there are fewer parts where Figs. 11 and 12 show the projection distances of the first the projection distances exceed the threshold than those in the principal element of a node in the conventional scheme “M-1,” other methods. This implies that we can obtain a lower number in the reference method “M-2,” and in the proposed method of false detections in the proposed method. Next, from Fig. 12, “M-3.” From these figures, we can see that, as a general where N = 100, similar to the case of N = 50, during the trend, the value of the projection distance increases during the attack period, we can see that the proposed method can detect time period of 2500–5000 s, when the attacks were executed. the anomaly. However, different from the case of N = 50, in In particular, in the proposed method “M-3,” the value of case of “M-1,” the projection distances increase. the projection distance rapidly increases at 2500 s and then This is because, in case of N = 100, there are significant sharply decreases at 5000 s as well. On the contrary, for the changes in the network environment. This causes the predefined conventional method “M-1,” the large projection distances can baseline profile and the present network state to dramatically be found through the whole period, and they do not descend at differ. Therefore, compared with the case of N = 50 during the time when attacks stop. This is the reason behind the lower the normal period, we can see that “M-1” produces more false detection rates (DRs) and a large number of false positives detections. Meanwhile, compared with the case of N = 50 in of the conventional method “M-1.” For the method “M-2,” “M-3,” we can see that although there is a number of parts comparing with “M-3”, the values of the projection distance that exceed the threshold, it generates less false detection than are relatively small and are scattered in a wide range; the result the other methods. From these facts, we can see that “M-1” is better than “M-1” and worse in contrast with “M-3.” cannot adapt with the changing environment. “M-1” does not As evident from Fig. 11, where the value of N is 50 during reflect the whole network state since it only represents the the attack period, “M-3” yields a higher projection distance and temporary state of the network. For “M-2,” by updating the can detect the current anomaly compared to the other methods. training data set, it can adapt to the changing environment to During the normal period, it should be noted that there are some extent. Note that, in “M-3,” the false detection in the case many points that exceed the threshold in “M-1” and “M-2.” of N = 100 is higher than that of N = 50. This is caused by On the other hand, there is a significant peak at around 9000 s, the increase in the RERR packets when the link is disconnected and this peak largely exceeds the threshold in “M-3.” Despite due to the network mobility and size, in addition to the sudden being a normal period when there is no attack, because of the increase of the RREQ packets when a large number of routeAuthorized licensed use limited to: Arulmigu Kalasalingam College of Engineering. Downloaded on August 04,2010 at 09:46:45 UTC from IEEE Xplore. Restrictions apply.
  9. 9. NAKAYAMA et al.: DYNAMIC ANOMALY DETECTION SCHEME FOR AODV-BASED MOBILE AD HOC NETWORKS 2479 TABLE I PERFORMANCE OF DETECTION MODULE BY PROJECTION DISTANCE: N = 50 TABLE II PERFORMANCE OF DETECTION MODULE BY PROJECTION DISTANCE: N = 100 requests occur at the same time. These trends become more (Pentium 4, 2.4 GHz), this computation time of “M-3” is apparent as the scale of the network becomes larger. Same below 10 ms. Considering the importance of network security reasons can be found for the increase of the false positives. and the increasing power of ad hoc nodes, we believe that Increasing the dimensionality may be effective to reduce these our proposed method can be a possible choice for performing false detections. anomaly detection in MANETs. In greater detail, Tables I and II display the average of the DR and the false positive rate (FPR) about the projection distance, VI. C ONCLUSION respectively. Based on the results shown in these two tables, we can see that the proposed method “M-3” provides the highest In this paper, a new dynamic anomaly detection system average DR and the lowest FPR. Compared to the conventional for MANETs has been proposed. For enhancing the security method “M-1,” the proposed method “M-3” increases the av- in MANETs, which are vulnerable to attacks, robust learning erage DR by more than 25%. In addition, the average FPR methods against these attacks are required. To differentiate an is decreased by more than 10% in “M-1.” Furthermore, as an attack state from the normal state, we have defined multidi- effectiveness of dynamic learning, “M-3” increases the average mensional features based on the characteristics of these attacks DR by more than 15% compared to “M-2” in our simulation. and utilized the projection distance using PCA based on sta- On the other hand, the FPR of “M-3” is only marginally better tistical theory. Our proposed system demonstrates an effective than that of “M-2.” This is because we approximately fixed the performance in terms of high DRs and low FPRs against five FPR in the range of 10%–20% for clearly understanding the simulated attacks, in addition to the scalability of the proposed difference of the DR between “M-2” and “M-3.” scheme clarified by the simulation results obtained from two Now, we evaluate the computational complexities of these distinct network topologies of varying sizes. three methods. The computational complexity of “M-1” using Future works will be focused on the various routing proto- the initial data is the lowest, and its order is O(1). “M-2” and cols in the MANET architecture. Although AODV is a major “M-3” compute the principal elements using PCA, which needs routing protocol in MANETs, new protocols are emerging, to compute the mean and the covariance. The order is O(p2 ) e.g., dynamic MANET on-demand protocol (DYMO) [43]. for computing both the mean and covariance, where p is the We will evaluate these protocols and give an analysis for the number of features described in Section IV-A. The order of additional types of attacks to further improve the accuracy finding the first principal element depends on the technique of the overall system. Moreover, in [44] and [45], Yan et al. of finding the first eigenvalue end eigenvector. For example, reported an interesting scheme with the context of studies on the the complexity of the power method [42] is O(k × p2 ) for intrusion detection system (IDS). The proposed IDS autonomic finding the first k eigenvalues and eigenvectors. The “M-2” event analysis system that is represented by description logics is O(p2 ), where k = 1. Because “M-3” has mi times more allows inferring the attack scenarios and enabling the attack learning processes than those of “M-2,” the complexity of knowledge semantic queries. To cite a case, first, using our “M-3” is O(mi × p2 ) in total. Apparently, there is a tradeoff proposed system to detect attacks and then rigorously applying between the computational complexity and the accuracy of the this IDS to analyze these attacks may bring about a reliable DRs. We actually tested the computation time of our proposed approach. Our future works will comprise of feasibility studies method by considering p = 14. Using a Linux-based computer on these more intelligent detection schemes in MANETs.Authorized licensed use limited to: Arulmigu Kalasalingam College of Engineering. Downloaded on August 04,2010 at 09:46:45 UTC from IEEE Xplore. Restrictions apply.
  10. 10. 2480 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 5, JUNE 2009 R EFERENCES [32] G. Vigna, S. Gwalani, K. Srinivasan, E. Belding-Royer, and R. Kemmerer, “An intrusion detection tool for AODV-based ad hoc wireless networks,” [1] P. Argyroudis and D. O’Mahony, “Secure routing for mobile ad hoc in Proc. 20th ACSAC, Dec. 2004, pp. 16–27. networks,” Commun. Surveys Tuts., vol. 7, no. 3, pp. 2–21, Third Quarter, [33] C. Tseng, P. Balasubramanyam, C. Ko, R. Limprasittiporn, J. Rowe, and 2005. K. Levitt, “A specification-based intrusion detection system for AODV,” [2] L. Zhou and Z. Haas, “Securing ad hoc networks,” IEEE Netw., vol. 13, in Proc. 1st ACM Workshop SASN, Oct. 2003, pp. 125–134. no. 6, pp. 24–30, Nov./Dec. 1999. [34] Y. Huang, W. Fan, W. Lee, and P. Yu, “Cross-feature analysis for detecting [3] M. Zapata and N. Asokan, “Securing ad hoc routing protocols,” in Proc. ad-hoc routing anomalies,” in Proc. 23rd ICDCS, May 2003, pp. 478–487. 3rd ACM Workshop WiSE, Sep. 2002, pp. 1–10. [35] Y. Huang and W. Lee, “Attack analysis and detection for ad hoc routing [4] P. Papadimitratos and Z. Haas, “Secure data transmission in mobile ad hoc protocols,” in Proc. 7th Int. Symp. RAID, Sep. 2004, pp. 125–145. networks,” in Proc. ACM Workshop WiSE, Sep. 2003, pp. 41–50. [36] B. Sun, K. Wu, and U. Pooch, “Towards adaptive intrusion detection [5] W. Diffie and M. E. Hellman, “New directions in cryptography,” IEEE in mobile ad hoc networks,” in Proc. IEEE Global Telecommun. Conf. Trans. Inf. Theory, vol. IT-22, no. 6, pp. 644–654, Nov. 1976. GLOBECOM, Nov./Dec. 2004, pp. 3551–3555. [6] A. J. Menezes, S. A. Vanstone, and P. C. V. Oorschot, Handbook of [37] R. Duda, P. Hart, and D. Stork, Pattern Classification and Scene Analysis. Applied Cryptography. Boca Raton, FL: CRC, 1996. New York: Wiley, 1973. [7] J. Edney and W. Arbaugh, Real 802.11 Security: Wi-Fi Protected Access [38] J. Tsumochi, K. Masayama, H. Uehara, and M. Yokoyama, “Impact of 802.11i. Upper Saddle River, NJ: Pearson, 2004. mobility metric on routing protocols for mobile ad hoc networks,” in Proc. [8] C.-K. Toh, Ad Hoc Mobile Wireless Networks—Protocol and Systems. IEEE Pacific Rim Conf. Commun., Comput. Signal Process. PACRIM, Upper Saddle River, NJ: Pearson, 2002. Aug. 2003, pp. 322–325. [9] A. Mishra, K. Nadkarni, and A. Patcha, “Intrusion detection in wireless [39] D. Maltz, J. Broch, J. Jetcheva, and D. Johnson, “The effects of on- ad hoc networks,” IEEE Wireless Commun., vol. 11, no. 1, pp. 48–60, demand behavior in routing protocols for multihop wireless ad hoc net- Feb. 2004. works,” IEEE J. Sel. Areas Commun., vol. 17, no. 8, pp. 1439–1453, [10] Y. Waizumi, Y. Sato, and Y. Nemoto, “A network-based anomaly detection Aug. 1999. system using multiple network features,” in Proc. 3rd Int. Conf. WEBIST, [40] C. Perkins, E. Royer, S. Das, and M. Marina, “Performance comparison Mar. 2007, pp. 410–413. of two on-demand routing protocols for ad hoc networks,” IEEE Pers. [11] H. Ebbinghaus, Memory: A Contribution to Experimental Psychology. Commun., vol. 8, no. 1, pp. 16–28, Feb. 2001. New York: Teachers College, 1913. [41] T. Camp, J. Boleng, and V. Davies, “A survey of mobility models for [12] I. London, “An ideal equation derived for a class of forgetting curves,” ad hoc network research,” Wirel. Commun. Mob. Comput., vol. 2, no. 5, Psychol. Rev., vol. 57, no. 5, pp. 295–302, Sep. 1950. pp. 483–502, Sep. 2002. [13] P. Ning and K. Sun, “How to misuse AODV: A case study of insider [42] G. Golub and C. V. Loan, “Matrix computations,” in Johns Hopkins Stud- attacks against mobile ad-hoc routing protocols,” in Proc. 4th Annu. IEEE ies in Mathematical Sciences, 3rd ed. Baltimore, MD: Johns Hopkins Inf. Assurance Workshop, Jun. 2003, pp. 60–67. Univ. Press, 1996. [14] W. Wang, Y. Lu, and B. Bhargava, “On vulnerability and protection of ad [43] I. Chakeres and C. Perkins, Dynamic MANET on-demand (DYMO) hoc on-demand distance vector protocol,” in Proc. 10th ICT, Feb. 2003, routing, Jul. 2007. IETF Internet Draft, draft-ietf-manet-dymo-10.txt. pp. 375–382. [44] W. Yan, E. Hou, and N. Ansari, “Description logics for an autonomic ids [15] M. Hollick, J. Schmitt, C. Seipl, and R. Steinmetz, “On the effect of node event analysis system,” Comput. Commun., vol. 29, no. 15, pp. 2841– misbehavior in ad hoc networks,” in Proc. IEEE Global Telecommun. 2852, Sep. 2006. Conf. GLOBECOM, Jun. 2004, pp. 3759–3763. [45] W. Yan, E. Hou, and N. Ansari, “Extracting and querying network attack [16] C. Perkins, E. Belding-Royer, and S. Das, Ad hoc On-Demand Distance scenarios knowledge in IDS using PCTCG and alert semantic networks,” Vector (AODV) Routing, Jul. 2003. IETF RFC 3561 (Experimental). in Proc. IEEE ICC, May 2005, pp. 1512–1517. [17] Network Simulator—NS (ver. 2). [Online]. Available: http://nsnam.isi. edu/nsnam/ [18] M. Zapata, Secure ad hoc on-demand distance vector (SAODV) routing, Sep. 2006. IETF Internet Draft, draft-guerrero-manet-saodv-06.txt. [19] A-SAODV Homepage. [Online]. Available: http://saodv.cefriel.it/ [20] D. Cerri and A. Ghioni, “Securing AODV: the A-SAODV secure rout- Hidehisa Nakayama (M’05) received the B.E., ing prototype,” IEEE Commun. Mag., vol. 46, no. 2, pp. 120–125, M.S., and Ph.D. degrees in information sciences Feb. 2008. from Tohoku University, Sendai, Japan, in 2000, [21] R. Rivest, A. Shamir, and L. Adleman, “A method for obtaining digital 2002, and 2005, respectively. signatures and public key cryptosystems,” Commun. ACM, vol. 21, no. 2, He is currently a Senior Assistant Professor with pp. 120–126, Feb. 1978. Tohoku Institute of Technology, Sendai. He has been [22] D. Eastlake, III and P. Jones, US Secure Hash Algorithm 1 (SHA1), engaged in research on intelligent sensor technology, Sep. 2001. IETF RFC 3174 (Informational). wireless mobile ad hoc networks, computer network- [23] H. Yih-Chun and A. Perrig, “A survey of secure wireless ad hoc routing,” ing, character string analysis, pattern recognition, IEEE Security Privacy, vol. 2, no. 3, pp. 28–39, May/Jun. 2004. and image processing. [24] H. Yih-Chun, A. Perrig, and D. Johnson, “Ariadne: A secure on-demand Dr. Nakayama is a member of the IEEE Commu- routing protocol for ad hoc networks,” Wirel. Netw., vol. 11, no. 1/2, nications Society, the Institute of Electronics, Information, and Communication pp. 21–38, Jan. 2005. Engineers (IEICE), and the Information Processing Society of Japan (IPSJ). He [25] H. Yih-Chun, D. Johnson, and A. Perrig, “SEAD: Secure efficient distance received the Paper Award for Young Researchers from the IPSJ Tohoku Chapter vector routing for mobile wireless ad hoc networks,” Ad Hoc Netw., vol. 1, in 2000 and the Best Paper of Pattern Recognition Award at SCI 2003. no. 1, pp. 175–192, Jul. 2003. [26] M. Bechler, H.-J. Hof, D. Kraft, F. Pahlke, and L. Wolf, “A clusterbased security architecture for ad hoc networks,” in Proc. 23rd Annu. Joint Conf. IEEE Comput. Commun. Soc. INFOCOM, Mar. 2004, pp. 2393–2403. [27] M. Ramkumar and N. Memon, “An efficient key predistribution scheme for ad hoc network security,” IEEE J. Sel. Areas Commun., vol. 23, no. 3, Satoshi Kurosawa received the B.E. degree from pp. 611–621, Mar. 2005. Miyagi University of Education, Sendai, Japan, in [28] K. Sanzgiri, D. LaFlamme, B. Dahill, B. N. Levine, C. Shields, and 2004 and the M.S. degree from Tohoku University, E. M. Belding-Royer, “Authenticated routing for ad hoc networks,” IEEE Sendai, in 2006. J. Sel. Areas Commun., vol. 23, no. 3, pp. 598–610, Mar. 2005. He is currently with the Information Technol- [29] H. Deng, W. Li, and D. Agrawal, “Routing security in ad hoc networks,” ogy R&D Center, Mitsubishi Electric Corporation, IEEE Commun. Mag., vol. 40, no. 10, pp. 70–75, Oct. 2002. Kamakura, Japan. His recent work has focused on ad [30] S. Lee, B. Han, and M. Shin, “Robust routing in wireless ad hoc net- hoc routing protocols and sensor network security. works,” in Proc. 31st ICPP Workshops, Aug. 2002, pp. 73–78. His research interests lie in the field of wireless [31] O. Kachirski and R. Guha, “Effective intrusion detection using multi- networking, particularly ad hoc network security. ple sensors in wireless ad hoc networks,” in Proc. 36th Annu. HICSS, Mr. Kurosawa received the Dean of the Graduate Jan. 2003, pp. 57–64. School of Information Sciences Award in 2005.Authorized licensed use limited to: Arulmigu Kalasalingam College of Engineering. Downloaded on August 04,2010 at 09:46:45 UTC from IEEE Xplore. Restrictions apply.
  11. 11. NAKAYAMA et al.: DYNAMIC ANOMALY DETECTION SCHEME FOR AODV-BASED MOBILE AD HOC NETWORKS 2481 Abbas Jamalipour (S’90–M’96–SM’00–F’07) re- Nei Kato (M’03–A’04–SM’05) received the M.S. ceived the Ph.D. degree from Nagoya University, and Ph.D. degrees from Tohoku University, Sendai, Nagoya, Japan. Japan, in 1988 and 1991, respectively. He is currently with the School of Electrical Infor- Since 1991, he has been with Tohoku Univer- mation Engineering, University of Sydney, Sydney, sity, where he is currently a Full Professor with Australia. He is the author of the first book on the Graduate School of Information Sciences. He wireless IP, as well as two other books, and has has published more than 120 papers in journals and coauthored five books and over 180 technical papers, peer-reviewed conference proceedings. He has been all in the field of mobile communications networks. engaged in research on computer networking, wire- He is also the author of several invited papers. His less mobile communications, image processing, and areas of research are wireless data communication neural networks. networks, wireless IP networks, next-generation mobile networks, traffic con- Dr. Kato is a member of the Institute of Electronics, Information, and trol, network security and management, and satellite systems. He was one Communication Engineers (IEICE). He has served as a Symposium Cochair of the first researchers to disseminate the fundamental concepts of next- at GLOBECOM’07 and ChinaCom’08 and as a Technical Program Committee generation mobile networks and broadband convergence networks, as well as member on a large number of IEEE international conferences, including the integration of wireless local area networks and cellular networks, some of ICC, GLOBECOM, WCNC, and HPSR. He was a Co-Guest Editor for the which are being gradually deployed by industry and included in the ITU-T Journal of Communications and Networks (JCN) Special Issue on Broadband standards. Convergence Networks (BcNs) in 2005. Since 2006, he has been the Technical Dr. Jamalipour is a Fellow of the Institute of Engineers Australia, Editor of IEEE WIRELESS COMMUNICATIONS. Since 2008, he has been an IEEE Distinguished Lecturer, the Editor-in-Chief of IEEE WIRELESS the Editor of the IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS. COMMUNICATIONS, and a Technical Editor of several scholarly journals, He is the corecipient of the 2005 Distinguished Contributions to Satellite including IEEE COMMUNICATIONS, the Wiley International Journal of Com- Communications Award from the IEEE Communications Society, Satellite and munication Systems, Journal of Communication Networks, etc. He has been a Space Communications Technical Committee, the corecipient of the FUNAI keynote speaker at many prestigious conferences. He served as the Chair of the Information Science Award in 2007, and the corecipient of the 2008 TELCOM Satellite and Space Communications Technical Committee (TC) from 2004 to System Technology Award from the Foundation for Electrical Communications 2006 and is currently the Vice Chair of Communications Switching and Routing Diffusion. He is serving as an expert member of the Telecommunications TC and the Chair of Chapters Coordinating Committee, Asia-Pacific Board, all Council, Ministry of Internal Affairs and Communications, Japan. with the IEEE Communications Society. He is a voting member of the IEEE GITC and IEEE WCNC Steering Committees. He was the Vice Chair of the IEEE WCNC from 2003 to 2006, the Program Chair of SPECTS2004, the Chair of symposiums at IEEE GLOBECOM 2005 to 2007 and IEEE ICC 2005 to 2008, as well as many other conferences. He has received several prestigious awards, such as the 2005 Telstra Award for Excellence in Teaching, the 2006 IEEE Communications Society Best Tutorial Paper Award, and the 2006 IEEE Distinguished Contribution to Satellite Communications Award. Yoshiaki Nemoto (S’72–M’73–SM’05) received the B.E., M.E., and Ph.D. degrees from Tohoku Uni- versity, Sendai, Japan, in 1968, 1970, and 1973, respectively. He is a Full Professor with the Graduate School of Information Sciences, Tohoku University, where he has also been an Executive Vice President since 2008. He has been engaged in research on microwave networks, communication systems, computer net- work systems, image processing, and handwritten character recognition. Dr. Nemoto is a Fellow of the Institute of Electrical, Information, and Communication Engineers (IEICE) and the Information Processing Society of Japan (IPSJ). He is a corecipient of the 1982 Microwave Prize from the IEEE Microwave Theory and Techniques Society, the 2005 Distinguished Contri- butions to Satellite Communications Award from the IEEE Communications Society, and the FUNAI information Science Award.Authorized licensed use limited to: Arulmigu Kalasalingam College of Engineering. Downloaded on August 04,2010 at 09:46:45 UTC from IEEE Xplore. Restrictions apply.